Running model on PDFs and Generating tokens for words.json from PDF

[Nikhilsonawane07]

I am generating tokens for Table detection model from PDF using following script.

And then converting pdf to images. However I am getting an error while running inference pipeline.

Please help with the issue. Also if anyone can share there script to run model on pdfs, it will be great ! Thanks is advance!

[aostiles]

I ran into a similar problem and fixed it. In my case, it was because Rect() expects four params: Rect(x0, y0, x1, y1) where (x0, y0) is the bottom-left and (x1, y1) is the top-right.

I'm using easyocr. To represent a bounding box, it returns a list of four coordinate-pairs. So I had to grab the two relevant coordinates.

I also ran into an issue serializing int64s. Here's my OCR code:

import json
import easyocr

reader = easyocr.Reader(['en'])
result = reader.readtext('path/to/image.jpg')
words = []

for _, word in enumerate(result):
    bbox_raw = word[0]
    bbox = [bbox_raw[0][0], bbox_raw[0][1], bbox_raw[2][0], bbox_raw[2][1]]
    text = word[1]
    words.append({"text": text, "bbox": bbox})

with open("path/to/image_words.json", "w") as file:
    json.dump(words, file, default=int)

I ran as follows:

python inference.py --image_dir ../path/to/img/ --words_dir ../path/to/words/ --out_dir ../results --structure_config_path ./structure_config.json --structure_model_path ./TATR-v1.1-Pub-msft.pth --mode extract --csv --detection_config_path ./detection_config.json --detection_model_path ./pubtables1m_detection_detr_r18.pth --visualize

[linkstatic12]

I would advise to use pdftools which is available in R. This library can be used in python. The pdftools are much more accurate when it comes to pdf manipulation

[Nikhilsonawane07]

I ran into a similar problem and fixed it. In my case, it was because Rect() expects four params: Rect(x0, y0, x1, y1) where (x0, y0) is the bottom-left and (x1, y1) is the top-right.

I'm using easyocr. To represent a bounding box, it returns a list of four coordinate-pairs. So I had to grab the two relevant coordinates.

I also ran into an issue serializing int64s. Here's my OCR code:

import json
import easyocr

reader = easyocr.Reader(['en'])
result = reader.readtext('path/to/image.jpg')
words = []

for _, word in enumerate(result):
    bbox_raw = word[0]
    bbox = [bbox_raw[0][0], bbox_raw[0][1], bbox_raw[2][0], bbox_raw[2][1]]
    text = word[1]
    words.append({"text": text, "bbox": bbox})

with open("path/to/image_words.json", "w") as file:
    json.dump(words, file, default=int)

I ran as follows:

python inference.py --image_dir ../path/to/img/ --words_dir ../path/to/words/ --out_dir ../results --structure_config_path ./structure_config.json --structure_model_path ./TATR-v1.1-Pub-msft.pth --mode extract --csv --detection_config_path ./detection_config.json --detection_model_path ./pubtables1m_detection_detr_r18.pth --visualize

Hey thanks for the reply, but I am looking to read text from pdf only not from images

[linkstatic12]

you can convert the PDF pages to images.